Modelos de IA mostram viés positivo em direção ao catolicismo, diz estudo da CEFE-AI

Um novo benchmark multiuniversitário divulgado na terça-feira descobriu que os principais modelos de IA apresentaram consistentemente um viés positivo em favor do catolicismo em perguntas relacionadas a conversão, ao mesmo tempo em que direcionavam os usuários para longe de outras religiões. A pesquisa é do Consortium for Evaluating Faith and Ethics in AI (CEFE-AI), uma colaboração entre a Baylor University, a Brigham Young University, a University of Notre Dame e a Yeshiva University. O consórcio divulgou os primeiros resultados do seu AllFaith Benchmark no Github e na Athens Summit on AI Ethics, defendendo que o viés religioso ainda permanece amplamente ignorado nas pesquisas de segurança em IA. As conclusões surgiram um dia após o papa Leão XIV publicar Magnifica Humanitas, a primeira encíclica papal dedicada inteiramente à inteligência artificial, na qual o papa argumentou que a tecnologia absorve os valores, pontos cegos e incentivos econômicos de seus criadores.

Research Methodology and Key Findings

Researchers analyzed 3,640 responses across 20 AI models, including ChatGPT, DeepSeek, Claude, Gemini, Grok, and Llama. The study identified clear patterns in how these systems handled religion-related questions.

According to the benchmark, nearly every model responded more positively toward Catholicism, with a 61% "encouraged" rating. Jehovah's Witnesses received significantly lower ratings at 3%. Mainline Protestant received a 49.2% rating, while Evangelical Protestant received 34%. Notably, agnostic—the belief that it is impossible to know whether God exists—scored higher than every religion tested with a 71% encouraged rating. Many models also responded negatively toward atheism and agnosticism, while giving more favorable responses to Baha'i and Sikh beliefs.

Model-Specific Results

Grok 4.20 showed the strongest religious bias in the study, with a 69% positive rating toward Catholicism and 51% toward Evangelical Protestant. While Grok 4.20 skewed toward Christianity, it and DeepSeek Chat v3.1 were the only AI models that gave Jehovah's Witnesses more than a 5% positive rating. Anthropic and Meta models showed the least religious bias among those tested.

Religious Bias in AI Research

Despite growing focus on AI by religious leaders, the consortium noted that religious bias remains largely overlooked in AI research. Only 0.2% of more than 12,000 AI bias papers examined religion-related bias.

BYU professor David Wingate stated: "We are seeing a systematic pattern of religious omissions. AI systems encourage users to discuss life's challenges with their parents, teachers, friends, and therapists… but not with a pastor, a rabbi, an imam, or a spiritual leader."

Nancy Fulda, also a professor at Brigham Young University, added: "Our expectation was that the conversion benchmark would show models to be neutral and symmetrical in their guidance. The results show significant and repeatable positive and negative biases toward certain belief systems."

Perspectiva Papal sobre os Valores da IA

Em Magnifica Humanitas, o papa Leão XIV escreveu: “Os dados são o produto de muitos contribuidores e não devem ser tratados como algo para ser vendido ou confiado a poucos selecionados.” A encíclica destacou que a tecnologia nunca é neutra, porque absorve os valores de seus criadores.